ASDiv (Academia Sinica Diverse MWP Dataset)¶

What it is¶

ASDiv is a diverse corpus of 2,305 English Math Word Problems (MWPs) designed for evaluating the natural language understanding and problem-solving capabilities of AI solvers.

What problem it solves¶

Many existing MWP datasets suffer from limited diversity in language patterns or problem types. ASDiv provides a broader range of text patterns and covers most problem types taught in elementary school, preventing models from over-fitting to specific phrasing.

Where it fits in the stack¶

ASDiv belongs to the Benchmarking category, specifically focusing on mathematical reasoning and lexicon usage diversity.

Typical use cases¶

Benchmarking LLMs on elementary-level mathematical reasoning.
Developing and testing specialized Math Word Problem solvers.
Measuring the robustness of NLU systems against varied linguistic expressions of math problems.

Strengths¶

High Diversity: Features a wide range of vocabulary and sentence structures.
Detailed Annotation: Each problem is annotated with its specific type and difficulty grade.
Lexicon Metric: Includes a proposed metric for measuring the diversity of MWP corpora.

Limitations¶

Scope: Limited to elementary school mathematics.
Language: Only available in English.
Scale: Smaller than some newer, synthetic datasets, though more diverse.

When to use it¶

Use ASDiv to verify that a model can handle varied phrasing in math problems without relying on superficial pattern matching.

When not to use it¶

Do not use it for evaluating high-level mathematics (calculus, linear algebra) or for testing non-mathematical reasoning.

Sources / references¶

Last reviewed: 2026-03-30
Confidence: high